Abstract 5017: A multimodal deep learning approach to predict survival in never smoking lung cancer patients

Purpose: Lung cancer in never-smokers (LCINS) is the fifth most common cause of cancer deaths worldwide. Identifying subjects who need more aggressive treatment could greatly help improve survival rates. In this study, we investigated whether a multimodal deep learning approach could be used to predict overall survival in a large cohort of never smoking lung adenocarcinoma patients from the Sherlock-Lung study. We leveraged multiple datasets containing different modalities, including clinical, digital pathology (WSIs: whole slide images), and genomic (driver gene mutations in EGFR, KRAS, TP53, RBM10, as well as CDKN2A deletions, ALK fusions, and mutation patterns like Kataegis) data. Method: The sample set contained 495 LCINS cases which were randomly split into training (60%), validation (15%), and testing (25%) sets. Tiles of 256 x 256 pixels were extracted from the WSIs and processed through a customized VGG-19 convolutional neural network. The output of this network was high-dimensional feature vectors representing morphological patterns in the tiles. Genomic and clinical data were modeled separately and then concatenated with the output of the CNN to generate a shared feature space, which was then modeled using a simple densely connected neural network. The final feature vector was then used to predict the risk score using Cox partial likelihood as the loss function. The risk predictor was then validated on the held-out test set. The above procedure was conducted with and without genomic data. We also conducted a gene-specific survival study across LCINS with different driver genes as well as restricting the analysis to stage I cases only. Results: We achieved an index of 0.86 for the concordance between predicted mortality risk and actual survival time on the test set using both approaches, with and without genomic data. The Kaplan-Meier curves for low- and high-risk groups, with the cutoff for the risk score determined by the median value, showed distinct survival patterns - patients predicted to have low risk exhibited a higher survival probability compared to those with higher scores. The log-rank test yielded a p-value of ≤0.05 for overall and gene-specific survival, adjusting for stage and age. Heatmaps generated using the activation map of the image model highlighted non-tumor regions as low-risk regions and tumor regions as high-risk regions. Conclusion: The multimodal approach strongly predicted overall survival across all stages and within stage I in LCINS. Specific driver gene mutations do not significantly enhance survival prediction, highlighting the benefits of modeling on multimodal information over using only molecular data. Such approaches for LCINS prognostication could be used in the future to guide effective treatment modalities. Citation Format: Monjoy Saha, Tongwu Zhang, Praphulla M.S. Bhawsar, Wei Zhao, Jianxin Shi, Soo-Ryum Yang, Jonas S. Almeida, Maria Teresa Landi. A multimodal deep learning approach to predict survival in never smoking lung cancer patients [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 5017.